Helen and I want AI to help us analyse Meta ads, prepare new variants and spot weak tests. We do not want an unproven agent moving money or publishing campaigns without approval.
That boundary is the interesting part.
I have spent a large amount on Facebook advertising across Collab365 over the years. That is a first-person business record, not an independently audited figure. What I can say with confidence is that the slow loop hurt.
We would create several headlines and images, upload them, wait for results, inspect the metrics, carry the learning back into another tool and start again.
The AI could write variants. It could not automatically see what happened after the advert went live.
An API is not an autonomous media buyer
Meta's Marketing API provides programmatic access to advertising objects and campaign insights, subject to app review, permissions, access tokens and platform rules.
That means software can retrieve performance data and, with the right approved access, manage parts of advertising workflows.
It does not mean a general AI assistant can safely connect itself to any ad account and optimise spend.
The original article referred to "Meta's new MCP access" as if Meta had supplied a finished official assistant for our workflow. We did not have adequate first-party evidence for that claim, so it has gone.
MCP can be used by a separate tool as an adapter around an API. The reliability, permissions and safety of that tool still need checking.
The loop we actually want
Our useful version looks like this:
- Pull read-only campaign and conversion data.
- Normalise attribution windows and campaign names.
- Ask AI to identify patterns and propose explanations.
- Make the uncertainty visible, especially when the sample is small.
- Draft new creative variants tied to one hypothesis.
- Require a human to approve audiences, claims, budget and publication.
- Enforce account-level spending limits outside the model.
- Record who changed what and why.
The AI can shorten the analysis and drafting work. It cannot decide whether a claim is legal, whether a conversion is valuable or whether a noisy early result deserves more money.
Faster is not automatically better
A small team can create far more variants with AI. That also makes it easier to generate clutter, split a budget across too many tests and mistake chance for insight.
The advantage is not producing the most adverts. It is closing a controlled learning loop without losing the human judgement that protects the customer and the bank account.
We are still building that loop. This is a design direction from our own business, not a tested promise that AI will lower your advertising costs or improve return on ad spend.